[MPlayer-dev-eng] [PATCH] Saturation & PSWAPD bugfix in mp3lib/dct64_3dnow.c & mp3lib/dct64_k7.c

[Thomas Orgis]

Am Fri, 7 Jul 2006 15:24:33 +0200
schrieb Diego Biurrun <diego at biurrun.de>: 

> On Tue, Jun 27, 2006 at 04:37:52PM +0800, Zuxy Meng wrote:
> > 2006/6/27, Guillaume Poirier <gpoirier at mplayerhq.hu>:
> > >Diego Biurrun wrote:
> > >>
> > >> How would I test this on my K6-III+?  It seems that dct64_k7.c was
> > >> compiled all along...
> > >
> > >Yes, it's compiled, but the ASM routines in dct64_k7.c aren't used in K6.
> > >You'd need to adapt the part the code in mp3lib/sr1.c::MP3_Init() to
> > >select the 3dnow code.
> > >
[...]
> It seems to work on my K6-III.  I played an MP3 file without problems.

Sorry for digging this old thing out, but Diego, did you observe any
improvement by using the 3DNowExt compared to plain 3DNow on the
K6-III+ (I suppose your's is a "+" one)?
I tested the mpg123 svn where I included the mp3lib K7 code a bit on my
own K6-III+.
It works, but it seems like 3DNowExt is actually a bit slower than
normal 3DNow:

thomas at knecht:~/mpg123-svn$ time src/mpg123 --cpu 3dnow -q -t $one_mp3_album
decoder: 3DNow

real	2m8.520s
user	2m6.690s
sys	0m1.610s

thomas at knecht:~/mpg123-svn$ time src/mpg123 -q -t $one_mp3_album
decoder: 3DNowExt

real	2m21.548s
user	2m20.040s
sys	0m1.270s

Can you confirm this? Perhaps one should still select plain 3DNow even if K6-III+
supports the extended one... or the 3DNowExt routine could be improved for
non-athlons. I guess that the 3DNow extensions aren't necessarily bad in the K6-III+.
On an Athlon64 (32bit), I remember that 3DNowExt improved and catched up with
SSE (but don't have numbers now, as well as I don't remember if it was a bit faster
or slower).


Alrighty then,

Thomas.